Responsible Artificial Intelligence for Predicting Electricity Demand in Residential Communities

Abstract

Artificial intelligence has gained increasing relevance in residential electricity-demand forecasting due to the temporal variability, behavior-dependence, weather-responsiveness, and linear-ignorant characterizability of household electricity consumption. Alongside accuracy improvements, emerging high-capacity machine-learning approaches in residential contexts present proliferation challenges for privacy, transparency, uncertainty, model drift, data minimization, and human responsibility. In this work, we construct a community-based simulation-aided Responsibility-AI methodology for short-term electricity-demand forecasting with accuracy and governance attributes. We simulated 1,036,800 residential household-hour observations (at hourly resolution) for a synthetic community of 240 households over 180 days, and evaluated four forecasting architectures (persistence, conventional long short-term memory [LSTM], explainable LSTM, and a Responsibility-AI ensemble of recurrent modeling, gradient boosting, to reduce data input sets, to estimate uncertainty, to increase transparency, to track drift, to promote human escalation) using community-based simulated performance metrics.

The Responsibility-AI architecture performed best regarding community-average overall forecasting performance, with MAE0.278kWh, RMSE0.386kWh, MAPE7.9%, PeakΔE6.8%, and 90% PI-Coverage94.1%. Its modeled privacy-risk score was reduced to 28/100 from 72/100 under community LSTM. A secondary logistic modeling approach revealed that hot extreme temperature, missing data, and comparatively volatile loads increased the modeled  

Artificial intelligence has gained increasing relevance in residential electricity-demand forecasting due to the temporal variability, behavior-dependence, weather-responsiveness, and linear-ignorant characterizability of household electricity consumption. Alongside accuracy improvements, emerging high-capacity machine-learning approaches in residential contexts present proliferation challenges for privacy, transparency, uncertainty, model drift, data minimization, and human responsibility. In this work, we construct a community-based simulation-aided Responsibility-AI methodology for short-term electricity-demand forecasting with accuracy and governance attributes. We simulated 1,036,800 residential household-hour observations (at hourly resolution) for a synthetic community of 240 households over 180 days, and evaluated four forecasting architectures (persistence, conventional long short-term memory [LSTM], explainable LSTM, and a Responsibility-AI ensemble of recurrent modeling, gradient boosting, to reduce data input sets, to estimate uncertainty, to increase transparency, to track drift, to promote human escalation) using community-based simulated performance metrics.

The Responsibility-AI architecture performed best regarding community-average overall forecasting performance, with MAE0.278kWh, RMSE0.386kWh, MAPE7.9%, PeakΔE6.8%, and 90% PI-Coverage94.1%. Its modeled privacy-risk score was reduced to 28/100 from 72/100 under community LSTM. A secondary logistic modeling approach revealed that hot extreme temperature, missing data, and comparatively volatile loads increased the modeled

probability of high-risk forecasting events, while modeled uncertainty escalation, drift monitoring, and human review reduced the modeled probability of high-risk forecasting events. Thus, our community-based methodology affirms that forecasting accuracy is compatible with the principles of human-centered governance, which considers transparency, privacy proportionality, calibrated uncertainty, operational reliability, and accountable oversight along with accuracy